Enabling Trust, Accountability, and Routine Use of AI-Enabled - - PowerPoint PPT Presentation

Enabling Trust, Accountability, and Routine Use of AI-Enabled Healthcare

Richard Giordano, Reham Al Tamime, Peter West

Web Science Institute

The HEALTH-I Team

We are living longer! But, this means more chronic illness. Doctors are facing incre reasi asing workl kloa

• ad and a need for more perso

sona nali lised sed care.

Diabetes 422 million worldwide Almost 4x more than 1980 (Mathers 2006) Heart failure re 6.5 million in USA Predicted to rise 46% by 2030 (American Heart Association 2017)

Challenges facing healthcare

Typical clinical scenarios

Example scenario:

• Patient presents to a doctor with heart

palpitations (irregular heart rate)

• She has had palpitations for the past few weeks
• Doctor examines the patient and asks about their

medications and lifestyle

• Doctor then refers patient for an EKG
• Three weeks latest, EKG doesn’t reveal anything.
• Doctor concludes the palpitations were caused by

atrial fibrillation.

Where could AI help medicine?

AI could help doctors get a better understanding of patients For our patient, AI could help the doctor identify the cause of the patient’s palpitations. But where can we get the data?

Patient-Generated Data

Any kind of data which a patient has recorded using their own means.

Wearables Fitbit, Apple Watch Journals Hand-written and electronic Health products Blood pressure cuffs, weighing scales Smartphone apps Google Fit, Strava

Food tracking

Ming ZY., Chen J., Cao Y., Forde C., Ngo CW., Chua T.S. (2018) Food Photo Recognition for Dietary Tracking: System and Experiment. In: Schoeffmann K. et al. (eds) MultiMedia Modeling. MMM 2018. Lecture Notes in Computer Science, vol 10705. Springer, Cham

Apps like DietLens help people track their calorie intake by analysing photos

• f food.

We asked 13 clinicians about the future of healthcare

What do doctors think the future of healthcare look like?

• 1. Fill the gaps between visits
• 2. Contextualise clinical data
• 3. Greater patient participation

In the hospital of the 2050?

Demo: http://flamingtempura.github.io/pgd-view

In the hospital of the 2050?

Monday 10am: palpitations

• 2 hours sleep

Tuesday 12pm: palpitations

• 4 hours sleep

Thursday 2pm: palpitations

• 3 hours sleep

Poor sleep was leading to worse palpitations.

But how will doctors perceive AI?

Some doctors have resisted the idea of using AI in healthcare. Will patient privacy be upheld? Will patients trust it?

Understanding the privacy concerns

• f mobile health

apps users

Reham Al Tamime

Research Problem

▪ The privacy decisions become more complex. ▪ Lack of understanding of users’ privacy concerns about the data collection and sharing practices in different contexts. ▪ Users are still provided with traditional privacy management options.

14

Research Goals

▪ To understand the privacy concerns of mobile health app users:

• a. To understand the comfort level in

sharing data across various contexts.

• b. To investigate how crowdsourcing can

help to understand privacy preferences.

15

Theory of contextual integrity

▪ Contextual integrity ties adequate protection for privacy to norms of specific contexts [Helen Nissenbaum, 2004]. ▪ Information gathering and dissemination be appropriate to that context and obey the governing norms of distribution within it. ▪ Capturing and specifying context elements of data collection and sharing are of great importance.

16

Scenarios of contexts

17

Type of data Location Whether the data is shared with a third party Purpose of data sharing Retention

Zooniverse

▪ Crowdsourcing, by definition, is the use of humans (at scale) to complete computationally difficult or time consuming tasks. ▪ Crowdsourcing for citizen science has been used to help annotate scientific datasets, such as Hubble Telescope images. ▪ 432 scenarios in total -> 16 subsets -> 27 scenario each.

18

19

Trust in AI healthcare technology

▪ Build framework that provides crowd-based privacy support for sharing data in various contexts. ▪ Design AI technologies that take into consideration users’ privacy concerns.

20

Black Box Medicine

Richard Giordano

Personalized Medicine

One-Size-Fits-All works only partially Fail to respond to treatment 38% depression 40% asthma 75% cancer Use of massive datasets and machine learning to design treatments for individuals

algorithms + data structures = programs

• Niklaus Wirth (1976)

○ Algorithms and data structures are intimately related ○ If you want to sort, use arrays

• Opacity of algorithms

○ Trade secrets ○ Technical literacy ○ Characteristics of the algorithm and the scale required to apply them usefully ○ National Nurses Union ■ “Algorithms are simple mathematical formulas that nobody understands”

Algorithmic programming: (PL/I)

The Invisible Maniac was algorithmic

Machine learning

• Machines trained from data

○ Classifiers produce categories ○ Learners train on data (based on models) and produce weights ○ Inductive reasoning

• “Applications that cannot be programmed by hand”

Black Box Medicine

• Opaque computational models to make decisions or judgements related to health

care

• Use of large scale datasets and associated algorithms (and heuristics) to use

implicit, complex connections between multiple patient characteristics

• Algorithms are neither explicit nor transparent
• Relationships they capture cannot be explicitly understood
• Relationships often cannot be explicitly stated
• This is not deliberately hidden—it is a consequence of complexity

Challenges for the Patient

○

How to explain treatment plan to a patient

■ Patients often do not understand information or retain it ○

Subgroups have different levels of trust in healthcare

■ White women | African American women ■ Native born | Immigrants ○

Predictive categorization not based on who/what you are

■ A viewer likely to enjoy a movie (Netflix) ■ A customer likely to buy this item (amazon) ■ A teenager likely to commit a crime (NYC predictive policing) ■ A women likely to become pregnant ■ A genotype likely to respond to CBT to treat schizophrenia

Challenges for the Doctor

○

Treatments and care pathways rely on complex biological interactions through integration of of information from interdisciplinary fields holistically

■ Common in Systems Biology ■ Not part of mainstream clinical research

• How do we develop/train doctors?

○

Quality of machine learning relies on aspects of training data and models

■ Who is responsible? ○

Deductive reasoning in medicine

■ Test a theory empirically-–randomized clinical trials ○

Inductive reasoning (Black Box Medicine)

■ Pattern recognition ■ Inductive reasoning not trusted among medics since it yields false positives

Challenges

• Policy

○ Equity across populations

• Institutions

○ Governance structures

• Technical

○ Systems to produce an audit trail ■ (This is not trivial…)

Conclusion: Enabling Trust, Accountability, and Routine Use

• f AI-Enabled Healthcare

Policy and societal challenges relating to privacy, trust, and transparency (We can’t just throw programmers at the problem) Our challenge to the NExT++ Workshop:

How can we build the next generation of accountable AI?

Our Published Research

Al Tamime, Giordano R, Hall, W (2018) Observing Burstiness in Wikipedia Articles during New Disease Outbreaks Web Science e Conference, e, Amsterd rdam, Netherlan ands West P, Van Kleek M, Giordano R, and Weal M (2018) Common barriers to the use of patient- generated data across clinical settings. CHI 2018, Montreal, l, Canada West P, Van Kleek M, Giordano R, Weal M, and Shadbolt N (2017) Information quality challenges of patient-generated data in clinical practice. Frontiers rs Public c Health. West P, Giordano R, Van Kleek M, and Shadbolt N (2016) The quantified patient in the doctor’s

• ffice: Challenges and opportunities. CHI 2016,

6, San Jose, USA. (Honorable le Mention) See our posters outside!